Summary
The morning slot is dominated by two technical-but-real signals and a wall of business hype. The strongest is GLM-5.2: Z.ai shipped a frontier open-weight model (1M context, two reasoning-effort levels, MIT license, same price as 5.1), and @eliebakouch pulled out the one genuinely interesting engineering detail, an "IndexShare + KVShare" trick on the multi-token-prediction head that removes a training-inference mismatch and lifts speculative-decoding acceptance length 20%. The second is hardware: NVIDIA's Blackwell platform swept MLPerf Training 6.0, with CoreWeave training DeepSeek-V3 671B in two minutes on 8,192 Blackwell Ultra GPUs and Azure hitting Llama 3.1 405B in 7.07 minutes. Around those, the funding and acquisition feed is loud: SpaceX's $60B all-stock Cursor buy (Cursor's CEO @mntruell confirming he is joining), DeepSeek's first outside money at a $50B+ valuation, Grok 4.3 landing on Amazon Bedrock, and Mistral's Arthur Mensch teasing a new "fat but sparse" model family for July. Anthropic's @_sholtodouglas boosted Taste Labs, an $18.5M-seed startup trying to give models "taste" and end AI slop. The rest of the slot is noise: French political ranting, SpaceX stock euphoria, and off-topic crime threads, all skipped.
Posts
GLM-5.2 ships, with an MTP KV-sharing trick worth reading (@eliebakouch · blog · weights). Z.ai released GLM-5.2: frontier intelligence, open weights, a usable 1M-token context, two reasoning-effort levels (max and high), MIT license, same API price as GLM-5.1. The attached images are the substance. The first is a blog excerpt on "MTP with IndexShare and KVShare": GLM-5.2 improves its multi-token-prediction layer (the small draft head used for speculative decoding) with two goals, make the draft cheap and make speculative acceptance high. In multi-step MTP the indexer sits on the first step and its top-k indices are reused for later steps; with IndexShare the KV cache of a predicted hidden state pulls only from the target model's hidden states rather than mixing in the MTP layer's own, which eliminates a training-inference discrepancy that GLM-5.1 had. An ablation table shows acceptance length climbing from a 4.56 baseline to 5.10 with IndexShare+KVShare, 5.29 with rejection sampling, and 5.47 (+20%) with an end-to-end TV loss. The second image is an "agentic coding performance by effort level" chart: GLM-5.2 (max) reaches roughly 75% averaged over Terminal-Bench 2.1, DeepSWE, and SWE-Atlas QnA at about 83k output tokens per task, a large jump over GLM-5.1 (max,
58%) and closing on Claude Opus 4.8 (78%) though at higher token cost. This is the open-weight price war made concrete: a near-Opus coding model under MIT, with real speculative-decoding engineering behind the token efficiency.NVIDIA Blackwell sweeps MLPerf Training 6.0 (@nvidia · blog). The Blackwell platform posted the fastest time-to-train on every benchmark and the largest-scale run at 8,192 GPUs on GB200 NVL72 systems, and was the only platform with submissions across all seven benchmarks. The round added two new mixture-of-experts pretraining workloads. The attached image shows Azure's milestone (Llama 3.1 405B trained in 7.07 minutes), and a separate NVIDIA note clarifies GB300 NVL72 delivers 1.6x on DeepSeek-V3 pretraining at 512-GPU scale. This is the hardware supply side under the long-context, sparse-MoE models the research feed keeps shipping.
CoreWeave trained DeepSeek-V3 671B in two minutes (@eliebakouch reposting CoreWeave). All 671B parameters on 8,192 Blackwell Ultra GPUs over Spectrum-X Ethernet, the fastest recorded DeepSeek-V3 training run, set in the same MLPerf v6.0 round. Bakouch's quip ("can we get a new track for the nanogpt speedrun") is a joke, but the datapoint is the real one: frontier-scale pretraining wall-clock is collapsing.
DeepSeek raises $7.4B at a $50B+ valuation (@ns123abc · The Information). DeepSeek's first external funding, with an unusual structure: investors put money into an LP managed by CEO Liang Wenfeng (who wrote the biggest check, $2.8B), external investors get no voting rights and a five-year lockup, and only the Chinese national AI fund invests directly with voting rights and no lockup. A Chinese open-weight lab raising at frontier-lab scale while staying founder-controlled.
Cursor teases a from-scratch frontier model (@ns123abc). Reposting Nick Dobos: a new Cursor model is being teased, same size class as Claude Opus and GPT-5.5, trained from scratch (no more Kimi base), 10-20x more compute than Composer, pitched as generally intelligent rather than coding-only, shipping in "the next couple of weeks." Lands right as SpaceX closes the $60B Cursor acquisition.
Cursor's CEO confirms joining SpaceX (@mntruell). Michael Truell: "Excited to be joining forces with @SpaceX to build useful AI," responding to the SpaceX post that SpaceXAI has been jointly training a model with Cursor for release in Cursor and Grok Build. The acquisition is real per the principals, not just a rumor repost.
Grok 4.3 lands on Amazon Bedrock (@mattsgarman · AWS). AWS's CEO announces xAI's Grok 4.3 on Bedrock, xAI's first appearance as a Bedrock model provider. Grok 4.3 is reasoning-first with configurable effort (none/low/medium/high), and runs on Mantle, a new Bedrock inference engine built for price-performance. Another model added to the multi-provider enterprise menu.
Mistral teases a sparse model family (@eliebakouch reposting Arthur Mensch). Mensch: a new model is coming this summer, "fat indeed, but sparse," the start of a new family, with an early-access program in July for research, government, and industry partners. Bakouch's read ("no more dense model guys") frames it as Mistral moving to mixture-of-experts. No weights or benchmarks yet.
Taste Labs comes out of stealth to "end AI slop" (@_sholtodouglas). Anthropic's Sholto Douglas boosts Thais Castello Branco's Taste Labs, $18.5M seed co-led by CRV and Amplify, building a data and infrastructure layer to give models and agents "taste," framed as turning the fuzzy, subjective judgment problem into something trainable. A bet that the next hard problem after objective capability is aesthetic judgment.
SemiAnalysis RL Systems blog amplified (@eliebakouch · post). Bakouch boosts Kimbo Chen's SemiAnalysis piece on matching trainer and generator throughput in RL training (GRPO, PipelineRL, async RL, policy staleness, sandbox infra). The technical anchor for the day's RL-cost theme, also covered in the daily digest's Global View.
Claude Managed Agents launch post (@ClaudeDevs · blog). Anthropic's Applied AI team on getting agents into production, covering credentials, sandboxing, and observability, with case studies and getting-started tips. Production-agent infrastructure framing, light on technical novelty.
Anthropic's Sholto Douglas on humanoid force-feedback (@_sholtodouglas · UME). Boosting the Universal Manipulation Exoskeleton, a low-cost upper-limb rig with real-time haptic torque feedback for collecting force-aware teleoperation data. Off the core efficiency beat (robotics hardware), but a clean datapoint that compliant-manipulation data collection is getting cheap.
Robert Scoble's AI-news and YC-firehose posts (cluster of several, @Scobleizer). Scoble pushing alignednews.com (an AI-news aggregator built on the X API, ingesting ~30k posts/day), going through 193 new YC startups, a Midjourney hardware teaser for 6/17, and a newsletter on public trust in AI. Promotional and aggregational, no research substance. Skim.
Political and stock-hype noise (cluster, skip). @brivael (French culture-war and SpaceX/Elon euphoria), @spencerpratt (LA art-deco architecture rant), @WHFraudTF (DOL fraud recovery press releases), @AustinJustice (off-topic crime threads). No AI content.